Transcript PPT - IPCC at UO - University of Oregon

Parallel Performance Tools
Parallel Computing
CIS 410/510
Department of Computer and Information Science
Lecture 14 – Parallel Performance Tools
Performance and Debugging Tools
Performance Measurement
and Analysis:
–
–
–
–
–
–
–
–
–
Open|SpeedShop
HPCToolkit
Vampir
Scalasca
Periscope
mpiP
Paraver
PerfExpert
TAU
Introduction to Parallel Computing, University of Oregon, IPCC
Modeling and prediction
– Prophesy
– MuMMI
Debugging
– Stat
Autotuning Frameworks
– Active Harmony
Lecture 14 – Parallel Performance Tools
2
Performance Tools Matrix
TOOL
Scalasca
HPCToolkit
Vampir
Open|SpeedSho
p
Periscope
mpiP
Paraver
TAU
Profiling Tracing Instrumentation Sampling
X
X
X
X
X
X
X
X
X
X
X
Introduction to Parallel Computing, University of Oregon, IPCC
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Lecture 14 – Parallel Performance Tools
3
Open|SpeedShop
Krell Institute (USA)
http://www.openspeedshop.org
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
4
Open|SpeedShop Tool Set

Open Source Performance Analysis Tool Framework
❍
❍
❍
❍

Flexible and Easy to use
❍

User access through:
GUI, Command Line, Python Scripting, convenience scripts
Scalable Data Collection
❍
❍

Most common performance analysis steps all in one tool
Combines tracing and sampling techniques
Extensible by plugins for data collection and representation
Gathers and displays several types of performance information
Instrumentation of unmodified application binaries
New option for hierarchical online data aggregation
Supports a wide range of systems
❍
❍
Extensively used and tested on a variety of Linux clusters
Cray XT/XE/XK and Blue Gene L/P/Q support
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
5
Open|SpeedShop Workflow
srun –n4 –N1
osspcsamp
“srun
smg2000
–n4 –N1
–n smg2000
65 65 65 –n 65 65 65”
MPI Application
O|SS
Introduction to Parallel Computing, University of Oregon, IPCC
Post-mortem
Lecture 14 – Parallel Performance Tools
6
Central Concept: Experiments

Users pick experiments:
❍
❍

What to measure and from which sources?
How to select, view, and analyze the resulting data?
Two main classes:
❍
Statistical Sampling
◆periodically interrupt execution and record location
◆useful to get an overview
◆low and uniform overhead
❍
Event Tracing (DyninstAPI)
◆gather and store individual application events
◆provides detailed per event information
◆can lead to huge data volumes

O|SS can be extended with additional experiments
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
7
Performance Analysis in Parallel

How to deal with concurrency?
❍
Any experiment can be applied to parallel application
◆Important step: aggregation or selection of data
❍

O|SS supports MPI and threaded codes
❍
❍
❍
❍

Special experiments targeting parallelism/synchronization
Automatically applied to all tasks/threads
Default views aggregate across all tasks/threads
Data from individual tasks/threads available
Thread support (incl. OpenMP) based on POSIX threads
Specific parallel experiments (e.g., MPI)
❍
Wraps MPI calls and reports
◆MPI routine time
◆MPI routine parameter information
❍
The mpit experiment also store function arguments and return
code for each call
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
8
HPCToolkit
John Mellor-Crummey
Rice University (USA)
http://hpctoolkit.org
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
9
HPCToolkit
Integrated suite of tools for measurement and
analysis of program performance
 Works with multilingual, fully optimized
applications that are statically or dynamically
linked
 Sampling-based measurement methodology
 Serial, multiprocess, multithread applications

Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
10
HPCToolkit HPCToolkit Workflow
profile
execution
compile & link
call path
profile
[hpcrun]
source
code
optimized
binary
binary
analysis
program
structure
[hpcstruct]
• Performance Analysis through callpath sampling
– Designed for low overhead
– Hot path analysis
– Recovery of program structure from binary
presentation
[hpcviewer/
hpctraceviewer]
database
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
8
Image by John Mellor-Crummey
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
11
HPCToolkit DESIGN PRINCIPLES

Employ binary-level measurement and analysis
❍
❍

Use sampling-based measurement (avoid instrumentation)
❍
❍
❍

diagnosis typically requires more than one species of metric
Associate metrics with both static and dynamic context
❍

controllable overhead
minimize systematic error and avoid blind spots
enable data collection for large-scale parallelism
Collect and correlate multiple derived performance metrics
❍

observe fully optimized, dynamically linked executions
support multi-lingual codes with external binary-only libraries
loop nests, procedures, inlined code, calling context
Support top-down performance analysis
❍
natural approach that minimizes burden on developers
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
12
HPCToolkit Workflow
compile & link
profile
execution
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
[hpcstruct]
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
call stack
profile
program
structure
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
13
HPCToolkit Workflow
compile & link
profile
execution
call stack
profile
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
program
structure
[hpcstruct]

For dynamically-linked executables on stock Linux
❍

compile and link as you usually do: nothing special needed
For statically-linked executables (e.g. for Blue Gene, Cray)
❍
add monitoring by using hpclink as prefix to your link line
• uses “linker wrapping” to catch “control” operations
– process and thread creation, finalization, signals, ...
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
14
HPCToolkit Workflow
compile & link
profile
execution
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
[hpcstruct]

call stack
profile
program
structure
Measure execution unobtrusively
❍
launch optimized application binaries
• dynamically-linked applications: launch with hpcrun to measure
• statically-linked applications: measurement library added at link time
– control with environment variable settings
❍
collect statistical call path profiles of events of interest
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
15
HPCToolkit Workflow
compile & link
profile
execution
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
[hpcstruct]

call stack
profile
program
structure
Analyze binary with hpcstruct: recover program
structure
❍
analyze machine code, line map, debugging information
❍ extract loop nesting & identify inlined procedures
❍ map transformed loops and procedures to source
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
16
HPCToolkit Workflow
profile
execution
compile & link
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
[hpcstruct]

program
structure
Combine multiple profiles
❍

call stack
profile
multiple threads; multiple processes; multiple executions
Correlate metrics to static & dynamic program structure
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
17
HPCToolkit Workflow
compile & link
profile
execution
call stack
profile
[hpcrun]
optimize
d
binary
app.
source
binary
analysis
program
structure
[hpcstruct]

Presentation
❍
explore performance data from multiple perspectives
• rank order by metrics to focus on what’s important
• compute derived metrics to help gain insight
– e.g. scalability losses, waste, CPI, bandwidth
❍
❍
graph thread-level metrics for contexts
explore evolution of behavior over time
presentation
[hpcviewer/
hpctraceviewer]
Introduction to Parallel Computing, University of Oregon, IPCC
databas
e
interpret profile
correlate w/ source
[hpcprof/hpcprof-mpi]
Lecture 14 – Parallel Performance Tools
18
Analyzing
resultsResults
with hpcviewer
Analyzing
with hpcviewer
associated
source code
costs for
• inlined procedures
• loops
• function calls in full context
source pane
view control
metric display
navigation pane
Callpath to
hotspot
Introduction to Parallel Computing, University of Oregon, IPCC
metric pane
24
Image by John Mellor-Crummey
Lecture 14 – Parallel Performance Tools
19
Vampir
Wolfgang Nagel
ZIH, Technische Universität Dresden (Germany)
http://www.vampir.eu
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
20
Mission


Visualization of dynamics
of complex parallel processes
Requires two components
❍
❍
Monitor/Collector (Score-P)
Charts/Browser (Vampir)
Typical questions that Vampir helps to answer:
❍
❍
❍
What happens in my application execution during a given
time in a given process or thread?
How do the communication patterns of my application
execute on a real system?
Are there any imbalances in computation, I/O or memory
usage and how do they affect the parallel execution of my
application?
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
21
Event Trace Visualization with Vampir
• Alternative and supplement to automatic analysis
• Show dynamic run-time behavior graphically at any
level of detail
• Provide statistics and performance metrics
Timeline charts
– Show application activities and
communication along a time axis
Summary charts
– Provide quantitative results for the
currently selected time interval
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
22
Vampir – Visualization Modes (1)
Directly on front end or local machine
% vampir
CPU
CPU
CPU
CPU
Multi-Core
Program
CPU
CPU
CPU
Vampir 8
Score-P
CPU
Small/Medium sized trace
Introduction to Parallel Computing, University of Oregon, IPCC
Trace
File
(OTF2)
Thread parallel
Lecture 14 – Parallel Performance Tools
23
Vampir – Visualization Modes (2)
On local machine with remote VampirServer
% vampirserver start –n 12
% vampir
VampirServer
Vampir 8
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
Score-P
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
Many-Core
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
Program
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
CPU
Introduction to Parallel Computing, University of Oregon, IPCC
LAN/WAN
Trace
File
(OTF2)
Large Trace File
(stays on remote machine)
MPI parallel application
Lecture 14 – Parallel Performance Tools
24
Main Displays of Vampir


Timeline Charts:
❍
Master Timeline
❍
Process Timeline
❍
Counter Data Timeline
❍
Performance Radar
Summary Charts:
❍
Function Summary
❍
Message Summary
❍
Process Summary
❍
Communication Matrix View
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
25
Visualization of the NPB-MZ-MPI / BT trace
% vampir scorep_bt-mz_B_4x4_trace
Navigation Toolbar
Function Summary
Function Legend
Master Timeline
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
26
Visualization of the NPB-MZ-MPI / BT trace
Master Timeline
Detailed information about
functions, communication
and synchronization events
for collection of processes.
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
27
Visualization of the NPB-MZ-MPI / BT trace
Process Timeline
Detailed information about
different levels of function
calls in a stacked bar chart
for an individual process.
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
28
Visualization of the NPB-MZ-MPI / BT trace
Typical program phases
Initialisation Phase
Computation Phase
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
29
Visualization of the NPB-MZ-MPI / BT trace
Counter Data Timeline
Detailed counter
information over time for
an individual process.
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
30
Visualization of the NPB-MZ-MPI / BT trace
Performance Radar
Detailed counter
information over time for
a collection of processes.
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
31
Visualization of the NPB-MZ-MPI / BT trace
Zoom in: Computation Phase
MPI communication
results in lower
floating point
operations.
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
32
Vampir Summary

Vampir & VampirServer
❍
Interactive trace visualization and analysis
❍ Intuitive browsing and zooming
❍ Scalable to large trace data sizes (20 TByte)
❍ Scalable to high parallelism (200000 processes)
Vampir for Linux, Windows and Mac OS X
 Vampir does neither solve your problems
automatically nor point you directly at them
 Rather it gives you FULL insight into the
execution of your application

Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
33
Scalasca
Bernd Mohr and Felix Wolf
Jülich Supercomputing Centre (Germany)
German Research School for Simulation Sciences
http://www.scalasca.org
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
34

Scalable parallel performance-analysis toolset
❍

Focus on communication and synchronization
Integrated performance analysis process
❍
Callpath profiling
◆performance overview on callpath level
❍
Event tracing
◆in-depth study of application behavior

Supported programming models
❍
MPI-1, MPI-2 one-sided communication
❍ OpenMP (basic features)

Available for all major HPC platforms
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
35
Scalasca Project: Objective
Development of a scalable performance analysis
toolset for most popular parallel programming
paradigms
 Specifically targeting large-scale parallel
applications

100,000 – 1,000,000 processes / thread
❍ IBM BlueGene or Cray XT systems
❍

Latest release:
❍
Scalasca v2.0 with Score-P support (August 2013)
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
36
Scalasca: Automatic Trace Analysis
Idea
❍
❍
Automatic search for patterns of inefficient behavior
Classification of behavior and quantification of significance
Low-level
event trace
Analysis
High-level
result

Property

Call
path
Location
❍
❍
❍
Guaranteed to cover the entire event trace
Quicker than manual/visual trace analysis
Parallel replay analysis online
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
37
Measurement
library
HWC
Instr.
target
application
Optimized measurement configuration
Local event
traces
Parallel waitstate search
Summary
report
Wait-state
report
Report
manipulation
Scalasca Workflow
Scalasca trace analysis
Instrumented
executable
Which problem?
Where in the
program?
Which
process?
Instrumenter
compiler / linker
Source
modules
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
38
Callpath Profile: Computation
Execution
time excl.
MPI comm
Just 30% of
simulation
Widely
spread
in code
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
39
Callpath Profile: P2P Messaging
MPI pointto-point
communication time
P2P comm
66% of
simulation
Introduction to Parallel Computing, University of Oregon, IPCC
Primarily
in scatter
& gather
Lecture 14 – Parallel Performance Tools
40
Callpath Profile: P2P Synchronization
Point-topoint msgs
w/o data
Masses of
P2P sync.
operations
Processes
all equally
responsible
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
41
Scalasca Approach to Performance Dynamics
• Capture overview of performance dynamics via
time-series profiling
Overview • Time and count-based metrics
• Identify pivotal iterations - if reproducible
Focus
• In-depth analysis of these iterations via tracing
• Analysis of wait-state formation
Ne
w
In-depth
• Critical-path analysis
• Tracing restricted to iterations of interest
analysis
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
42
TAU Performance System®


Tuning and Analysis Utilities (20+ year project)
Performance problem solving framework for HPC
❍
❍

Integrated performance toolkit
❍
❍
❍
❍
❍

Integrated, scalable, flexible, portable
Target all parallel programming / execution paradigms
Multi-level performance instrumentation
Flexible and configurable performance measurement
Widely-ported performance profiling / tracing system
Performance data management and data mining
Open source (BSD-style license)
Broad use in complex software, systems, applications
http://tau.uoregon.edu
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
43
TAU History
1992-1995: Malony and Mohr work with Gannon on DARPA pC++ project work.
TAU is born. [parallel profiling, tracing, performance extrapolation]
1995-1998: Shende works on Ph.D. research on performance mapping. TAU v1.0
released. [multiple languages, source analysis, automatic instrumentation]
1998-2001: Significant effort in Fortran analysis and instrumentation, work with Mohr
on POMP, Kojak tracing integration, focus on automated performance analysis.
[performance diagnosis, source analysis, instrumentation]
2002-2005: Focus on profiling analysis tools, measurement scalability, and
perturbation compensation. [analysis, scalability, perturbation analysis, applications]
2005-2007: More emphasis on tool integration, usability, and data presentation. TAU
v2.0 released. [performance visualization, binary instrumentation, integration,
performance diagnosis and modeling]
2008-2011: Add performance database support, data mining, and rule-based analysis.
Develop measurement/analysis for heterogeneous systems. Core measurement
infrastructure integration (Score-P). [database, data mining, expert system,
heterogeneous measurement, infrastructure integration]
2012-present: Focus on exascale systems. Improve scalability. Add hybrid
measurement support, extend heterogeneous and mixed-mode, develop user-level
threading. Apply to petascale / exascale applications. [scale, autotuning, user-level]
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
44
General Target Computation Model in TAU

Node: physically distinct shared memory machine
❍
Message passing node interconnection network
Context: distinct virtual memory space within node
 Thread: execution threads (user/system) in context

Interconnection Network
Node
Node
physical
view
*
node memory
memory
VM
space
model
view
…
message
* Inter-node
communication
Node
SMP
memory
…
Context
Introduction to Parallel Computing, University of Oregon, IPCC
Threads
Lecture 14 – Parallel Performance Tools
45
TAU Architecture


TAU is a parallel performance framework and toolkit
Software architecture provides separation of concerns
❍
Instrumentation | Measurement | Analysis
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
46
TAU Observation Methodology and Workflow

TAU’s (primary) methodology for parallel performance
observation is based on the insertion of measurement probes
into application, library, and runtime system
❍
❍
❍

Performance experimentation workflow
❍
❍
❍
❍
❍

Code is instrumented to make visible certain events
Performance measurements occur when events are triggered
Known as probe-based (direct) measurement
Instrument application and other code components
Link / load TAU measurement library
Execute program to gather performance data
Analysis performance data with respect to events
Analyze multiple performance experiments
Extended TAU’s methodology and workflow to support
sampling-based techniques
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
47
TAU Components

Instrumentation
❍
❍
❍

Measurement
❍
❍
❍
❍
❍

Fortran, C, C++, OpenMP, Python, Java, UPC, Chapel
Source, compiler, library wrapping, binary rewriting
Automatic instrumentation
Internode: MPI, OpenSHMEM, ARMCI, PGAS, DMAPP
Intranode: Pthreads, OpenMP, hybrid, …
Heterogeneous: GPU, MIC, CUDA, OpenCL, OpenACC, …
Performance data (timing, counters) and metadata
Parallel profiling and tracing (with Score-P integration)
Analysis
❍
❍
❍
❍
Parallel profile analysis and visualization (ParaProf)
Performance data mining / machine learning (PerfExplorer)
Performance database technology (TAUdb)
Empirical autotuning
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
48
TAU Instrumentation Approach

Direct and indirect performance instrumentation
❍
❍

Support for standard program code events
❍
❍
❍

❍
❍

Routines, classes and templates
Statement-level blocks, loops
Interval events (start/stop)
Support for user-defined events
❍

Direct instrumentation of program (system) code (probes)
Indirect support via sampling or interrupts
Interval events specified by user
Atomic events (statistical measurement at a single point)
Context events (atomic events with calling path context)
Provides static events and dynamic events
Instrumentation optimization
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
49
TAU Instrumentation Mechanisms

Source code
❍
❍
Manual (TAU API, TAU component API)
Automatic (robust)
◆ C, C++, F77/90/95, OpenMP (POMP/OPARI), UPC
❍

Compiler (GNU, IBM, NAG, Intel, PGI, Pathscale, Cray, …)
Object code (library-level)
❍
Statically- and dynamically-linked wrapper libraries
◆ MPI, I/O, memory, …
❍

Executable code / runtime
❍
❍
❍
❍

Powerful library wrapping of external libraries without source
Runtime preloading and interception of library calls
Binary instrumentation (Dyninst, MAQAO, PEBIL)
Dynamic instrumentation (Dyninst)
OpenMP (runtime API, CollectorAPI, GOMP, OMPT)
Virtual machine, interpreter, and OS instrumentation
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
50
Instrumentation for Wrapping External Libraries

Preprocessor substitution
❍
❍
❍

Preloading a library at runtime
❍
❍

Header redefines a routine with macros (only C and C++)
Tool-defined header file with same name takes precedence
Original routine substituted by preprocessor callsite
Library preloaded in the address space of executing application
intercepts calls from a given library
Tool wrapper library defines routine, gets address of global
symbol (dlsym), internally calls measured routine
Linker-based substitution
❍
Wrapper library defines wrapper interface
◆wrapper interface then which calls routine
❍
Linker is passed option to substitute all references from
applications object code with tool wrappers
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
51
Automatic Source-level / Wrapper Instrumentation
PDT source
analyzer
Parsed
program
Application
source
tau_instrumentor
Instrumented
source
tau_wrap
Instrumentation
specification file
BEGIN_EXCLUDE_LIST
Foo
Bar
D#EMM
END_EXCLUDE_LIST
BEGIN_FILE_EXCLUDE_LIST
f*.f90
Foo?.cpp
END_FILE_EXCLUDE_LIST
BEGIN_FILE_INCLUDE_LIST
main.cpp
foo.f90
END_FILE_INCLUDE_LIST
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
52
MPI Wrapper Interposition Library

Uses standard MPI Profiling Interface
❍
Provides name shifted interface (weak bindings)
◆MPI_Send = PMPI_Send

Create TAU instrumented MPI library
❍
Interpose between MPI and TAU
◆-lmpi replaced by –lTauMpi –lpmpi –lmpi
❍

No change to the source code, just re-link application!
Can we interpose MPI for compiled applications?
❍
❍
Avoid re-compilation or re-linking
Requires shared library MPI
◆uses LD_PRELOAD for Linux
❍
❍
Approach will work with other shared libraries (see later slide)
Use TAU tau_exec (see later slide)
% mpirun -np 4 tau_exec a.out
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
53
Binary Instrumentation


TAU has been a long-time user of DyninstAPI
Using DyninstAPI’s binary re-writing capabilities,
created a binary re-writer tool for TAU (tau_run)
❍
❍
Supports TAU's performance instrumentation
Works with TAU instrumentation selection
◆files and routines based on exclude/include lists
❍
❍

TAU’s measurement library (DSO) is loaded by tau_run
Runtime (pre-execution) and binary re-writing supported
Simplifies code instrumentation and usage greatly!
% tau_run a.out –o a.inst
% mpirun –np 4 ./a.inst

Support PEBIL and MAQAO binary instrumentation
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
54
Library Interposition

Simplify TAU usage to assess performance properties
❍



Application, I/O, memory, communication
Designed a new tool that leverages runtime
instrumentation by pre-loading measurement libraries
Works on dynamic executables (default under Linux)
Substitutes routines (e.g., I/O, MPI, memory
allocation/deallocation) with instrumented calls
❍
❍
Interval events (e.g., time spent in write())
Atomic events (e.g., how much memory was allocated)
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
55
Library wrapping – tau_gen_wrapper

How to instrument an external library without source?
❍
❍

Build a library wrapper tools
❍
❍
❍

Source may not be available
Library may be too cumbersome to build (with instrumentation)
Used PDT to parse header files
Generate new header files with instrumentation files
Three methods: runtime preloading, linking, redirecting headers
Add to TAU_OPTIONS environment variable:
–optTauWrapFile=<wrapperdir>/link_options.tau

Wrapped library
❍
❍
❍
Redirects references at routine callsite to a wrapper call
Wrapper internally calls the original
Wrapper has TAU measurement code
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
56
TAU Measurement Approach

Portable and scalable parallel profiling solution
❍
❍
❍
❍

Portable and scalable parallel tracing solution
❍
❍



Multiple profiling types and options
Event selection and control (enabling/disabling, throttling)
Online profile access and sampling
Online performance profile overhead compensation
Trace translation to OTF, EPILOG, Paraver, and SLOG2
Trace streams (OTF) and hierarchical trace merging
Robust timing and hardware performance support
Multiple counters (hardware, user-defined, system)
Metadata (hardware/system, application, …)
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
57
TAU Measurement Mechanisms

Parallel profiling
❍
❍
❍
❍
❍
❍

Function-level, block-level, statement-level
Supports user-defined events and mapping events
Support for flat, callgraph/callpath, phase profiling
Support for parameter and context profiling
Support for tracking I/O and memory (library wrappers)
Parallel profile stored (dumped, shapshot) during execution
Tracing
❍
❍
❍
All profile-level events
Inter-process communication events
Inclusion of multiple counter data in traced events
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
58
Parallel Performance Profiling

Flat profiles
❍
❍

Callpath profiles (Calldepth profiles)
❍
❍
❍

Time spent along a calling path (edges in callgraph)
“main=> f1 => f2 => MPI_Send” (event name)
TAU_CALLPATH_DEPTH environment variable
Phase profiles
❍
❍
❍

Metric (e.g., time) spent in an event (callgraph nodes)
Exclusive/inclusive, # of calls, child calls
Flat profiles under a phase (nested phases are allowed)
Default “main” phase
Supports static or dynamic (per-iteration) phases
Parameter and context profiling
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
59
Performance Analysis


Analysis of parallel profile and trace measurement
Parallel profile analysis (ParaProf)
❍
❍



Performance data management (TAUdb)
Performance data mining (PerfExplorer)
Parallel trace analysis
❍
❍



Java-based analysis and visualization tool
Support for large-scale parallel profiles
Translation to VTF (V3.0), EPILOG, OTF formats
Integration with Vampir / Vampir Server (TU Dresden)
Integration with CUBE browser (Scalasca, UTK / FZJ)
Scalable runtime fault isolation with callstack debugging
Efficient parallel runtime bounds checking
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
60
Profile Analysis Framework
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
61
Performance Data Management (TAUdb)

Provide an open, flexible framework to support
common data management tasks
❍

Foster multi-experiment performance evaluation
Extensible toolkit to promote integration and reuse
across available performance tools
❍
❍
Supported multiple
profile formats:
TAU, CUBE, gprof,
mpiP, psrun, …
Supported DBMS:
PostgreSQL, MySQL,
Oracle, DB2,
Derby, H2
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
62
TAUdb Database Schema


Parallel performance profiles
Timer and counter measurements with 5 dimensions
❍
❍
❍
❍
❍

Physical location: process / thread
Static code location: function / loop / block / line
Dynamic location: current callpath and context (parameters)
Time context: iteration / snapshot / phase
Metric: time, HW counters, derived values
Measurement metadata
❍
❍
❍
Properties of the experiment
Anything from name:value pairs to nested, structured data
Single value for whole experiment or full context (tuple of
thread, timer, iteration, timestamp)
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
63
TAUdb Tool Support

ParaProf
❍
Parallel profile analyzer
◆visual pprof
❍
❍

2, 3+D visualizations
Single and comparative
experiment analysis
PerfExplorer
❍
Data mining framework
◆Clustering, correlation
❍
❍
❍
Multi-experiment analysis
Scripting engine
Expert system
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
64
ParaProf – Single Thread of Execution View
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
65
ParaProf – Full Profile / Comparative Views
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
66
How to explain performance?


Should not just redescribe the performance results
Should explain performance phenomena
❍
❍
❍

Need to add knowledge to do more intelligent things
❍
❍

What are the causes for performance observed?
What are the factors and how do they interrelate?
Performance analytics, forensics, and decision support
Automated analysis needs good informed feedback
Performance model generation requires interpretation
Build these capabilities into performance tools
❍
❍
❍
❍
Support broader experimentation methods and refinement
Access and correlate data from several sources
Automate performance data analysis / mining / learning
Include predictive features and experiment refinement
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
67
Role of Knowledge and Context in Analysis
Context Metadata
You have to
capture these...
Source
Code
Build
Environmen
t
Run
Environmen
t
Performanc
e Problems
Application
Machine
Execution
Performance Knowledge
...to understand
this
Introduction to Parallel Computing, University of Oregon, IPCC
Performance Result
Lecture 14 – Parallel Performance Tools
68
Score-P Architecture
TAU
Vampir
Scalasca
supplemental
instrumentation
+ measurement
support
Event traces (OTF2)
TAU
Periscope
Call-path profiles (CUBE4)
Online
interface
Hardware counter (PAPI)
Score-P measurement infrastructure
TAU
adaptor
Application (MPI, OpenMP, hybrid)
Instrumentation
Compiler
TAU
instrumentor
Introduction to Parallel Computing, University of Oregon, IPCC
OPARI 2
MPI wrapper
COBI
Lecture 14 – Parallel Performance Tools
69
For More Information …

TAU Website
http://tau.uoregon.edu
❍
Software
❍ Release notes
❍ Documentation

HPC Linux
http://www.hpclinux.com
Parallel Tools “LiveDVD”
❍ Boot up on your laptop or desktop
❍ Includes TAU and variety of other packages
❍ Include documentation and tutorial slides
❍
Introduction to Parallel Computing, University of Oregon, IPCC
Lecture 14 – Parallel Performance Tools
70